Face Recognition Algorithms Based On Subspaces Method

Face recognition is widely used in payment,video surveillance,and criminal investigation and so on.In consideration of robustness of Gabor,LBP and ASIFT to lighting change and deflection,and the little requirement of DBN network on the number of training samples,this master dissertation includes some improved face recognition algorithms and their experiments on the basis of the existing subspace algorithms.In order to solve existing subspace algorithms' over-reliance on the number of training samples,the improved face recognition algorithm,i.e."DBN face recognition algorithm",in which the DBN network is used as a classifier,is proposed,and three groups of experiments is designed to verify its effectiveness.Under the conditions of no lighting and pose change,the recognition rate of the algorithm is higher than that of the existing subspace algorithm.The experimental results show that the face recognition rate of DBN face recognition algorithm reaches 91.79% on ORL database,86.84% on Yale A database and 93.50% on CMU-PIE database respectively.In order to solve that it is difficult for "DBN face recognition algorithm" to resist lighting change and pose change,the improved face recognition algorithm based on 2DGabor and rotation & uniform invariant LBP,namely "2DGabor + LBP algorithm",is proposed.The 2DGabor is used to extract initial face features and to generate its Gabor images.Then the rotation & uniform invariant LBP operator is used to extract the features of Gabor images in order to achieve the robustness to lighting and pose change in face recognition.The experimental results show that the face recognition rate of the algorithm is 95.35% on ORL database,96.14% on Yale A database,and 95.3% on CMU-PIE database respectively.In order to solve that "DBN face recognition algorithm" and "2DGabor+LBP algorithm" have low face recognition rate on rotation and scale change images,the improved face recognition algorithm based on ASFIT and information entropy,namely "ASIFT + information entropy algorithm" is proposed.The 2D-maximum information entropy is used to screen the extracted ASIFT feature points and calculate the information entropy of feature points located in concentration regions,which is used as the secondary matching parameters to improve the recognition rate.The experimental results show that the face recognition rate of the algorithm is 95.2%and its matching time is 10.65 s on the CMU-PIE database.For pose change images,the face recognition rate reaches 96.85% and its matching time is 7.17 s.For rotated images,the face recognition rate reaches 96.7% and its matching time is 9.46 s.For scale-varying images,the face recognition rate reaches 95.7% and its matching time is 8.44 s.The existing work in this dissertation demonstrates that the "DBN face recognition algorithm","2DGabor+LBP algorithm" and "ASIFT + information entropy algorithm" are all applicable to face recognition on lesser training samples,and "2DGabor+LBP" algorithm is suitable to face recognition on lighting and pose change,and " ASIFT + information entropy algorithm " not only has the ability to resist lighting and pose change,but also improve its ability to deal with rotation and scale change.